Conclusion

Let’s get straight to the point. AI regulation will serve as a “tailwind for small and medium enterprises (SMEs).”

The U.S. government is considering the establishment of a “FINRA-like” oversight body to review AI models.

FINRA is a self-regulatory organization that oversees the U.S. securities industry. There is now talk of creating a similar structure for the AI industry. The idea is that a third party will assess the safety and fairness of AI models—sounds good, right? But who will bear the regulatory costs? This is where a structural “reversal” occurs.

The burden of review costs will fall on the large corporations that develop the models.
Meanwhile, SMEs that simply use the reviewed models via APIs will obtain “safety-certified” results without incurring costs.

This is not a coincidence. The very structure of the AI industry is generating this reversal.

The Value of AI Has Shifted from “Models” to “Infrastructure”

First, let’s clarify the premise. What is currently most valued in the AI industry?

It is not the models themselves. It is the infrastructure.

Databricks is valued at $188 billion (approximately ¥28 trillion). Companies providing data infrastructure are receiving higher valuations than many AI model development firms. OpenAI is said to be valued at $300 billion, but it is a composite of “models + infrastructure + products.” The existence of Databricks, which has reached this scale purely through infrastructure, clearly indicates what the market values.

Why is infrastructure important? The reason is simple.

Models become commoditized. When GPT-4 is released, within six months, Claude, Gemini, Llama, and Mistral will catch up. Performance differences continue to shrink. However, the data pipelines, learning infrastructure, and inference infrastructure required to run those models cannot be easily copied.

The “AI scorecard” proposed by OpenAI’s CFO, Sarah Fryer, also supports this trend. She highlighted two indicators:

• Cost per successful task
• Return per compute unit

Both of these metrics measure “infrastructure efficiency” rather than “model intelligence.” In other words, the evaluation criteria for AI investments are shifting from “how smart it is” to “how cheaply it can deliver results.”

Who Will Bear the Costs of Establishing the Oversight Body?

Now, let’s return to the discussion of a FINRA-like oversight body.

The annual operating budget for FINRA in the U.S. financial industry is about $1.6 billion (approximately ¥240 billion). This cost is borne by securities firms through membership fees and commissions. If we apply the same structure to AI, the companies that develop and provide models will be the ones subject to review.

OpenAI, Google, Meta, Anthropic, Microsoft—these companies will bear the review costs. While the exact scale is yet to be determined, based on the scale of FINRA, it would not be surprising if the industry as a whole incurs annual costs in the range of hundreds of millions to over a billion dollars.

Additionally, there are costs associated with establishing internal systems to comply with the reviews. Expanding compliance teams, documentation, and audit responses—large securities firms in the financial industry spend tens of millions of dollars annually just on compliance-related activities. AI companies will incur similar costs.

So, what about SMEs?

The answer is clear. Almost nothing will change.

SMEs do not develop models. They simply access APIs. Companies that use services like the ChatGPT API, Claude API, and Gemini API, which can be accessed for a few thousand to tens of thousands of yen per month, will not be subject to review.

Instead, they will be able to use the “certified” models at the same prices as before. If regulation ensures the safety of the models, the risk for SMEs will only decrease.

It’s Not “Free Riding”; It’s Just How the Structure Works

Describing this situation as “free riding” makes it sound like SMEs are being unfair. However, that is not the case.

This is an inevitability created by the structure of the AI industry itself.

The same thing happened with cloud computing. AWS, Azure, and GCP made massive capital investments to build data centers, obtain security certifications, and comply with regulations in various countries. SMEs built services on top of that infrastructure for a few thousand yen per month. No one calls this “free riding” because they are paying a fair price for the services.

The same is happening in AI. Large corporations develop models, bear the review costs, and provide them as APIs. SMEs pay a fair price in the form of API usage fees to utilize those results.

What differs is that the absence of regulatory costs functions as an “entry barrier.”

The hurdles for large corporations to enter AI model development are raised further by regulation. In addition to development costs, they must also account for review and compliance costs. As a result, model development will concentrate among a few giant companies, making the two-tier structure between these companies and SMEs using the models via APIs even more pronounced.

What SMEs Should Consider

Given this structure, what should SMEs do?

1. Compete on “how to use the API” rather than “which API to use”

Performance differences between models are shrinking. API prices continue to fall. OpenAI’s GPT-4o has seen API costs drop by about 90% compared to GPT-4 from a year ago. The input cost per million tokens has decreased from $30 for GPT-4 to $2.5 for GPT-4o. This trend will not stop.

In other words, “which AI to use” will no longer be a differentiating factor. What will set companies apart is “where and how to apply it in their business.” Local SMEs have contextual insights that large corporations do not. Only those familiar with the field can connect that context with AI.

2. Present regulation as a “reassurance” to customers

If an oversight body is established, SMEs can say, “The AI models we use have passed third-party reviews.” This can become a powerful sales tool, especially for B2B SMEs, as they will no longer need to prove safety themselves.

Previously, when saying, “We use AI,” they would be asked, “Is that safe?” With an oversight body, they can respond, “We use reviewed models.” This difference is significant.

3. Understand the cost reversal through numbers

Let’s do some calculations. Suppose a certain task costs ¥300,000 per month when done manually. Automating the same task with an AI API incurs an API usage fee of ¥5,000 per month. That’s 1/60th of the labor cost.

Even if regulation hypothetically doubles the API fee, it would still be ¥10,000 per month. That’s still 1/30th of the labor cost. Even if regulatory costs are ultimately passed on to API prices, the impact on SMEs would be minimal.

Rather, if regulation ensures the quality of the models, SMEs can achieve the best scenario of “affordable and safe.”

The Real Risk Is Not “Regulation” But “Doing Nothing”

Finally, I want to pose a question.

Is AI regulation really a risk for SMEs?

Regulatory costs will be borne by large corporations. API prices will continue to fall. The safety of the models will be guaranteed by third parties. There is no more favorable structure for SMEs than this.

The real risk is failing to recognize this structural change and doing nothing.

When competitors start using AI via APIs to run a ¥300,000 task for ¥5,000, what will your company do? Worrying about regulatory trends is not the priority.

While large corporations pay tens of millions of dollars in regulatory costs, SMEs can utilize those results for just a few thousand yen per month. This asymmetry is historically rare.

The same thing happened with cloud computing. Companies that delayed adoption due to concerns like “security worries” or “not understanding it well” ultimately lost in the cost competition. The same will happen with AI, and this time, the speed of change is even faster.

The first step is to try hitting the API in your own business. There’s no need for approval for an experiment that can start for just ¥5,000.